Extended integral wall-model for large-eddy simulations of ... › wp-content › uploads › 2018...
Transcript of Extended integral wall-model for large-eddy simulations of ... › wp-content › uploads › 2018...
Extended integral wall-model for large-eddy simulations of compressible wall-bounded1
turbulent flows2
M. Catchirayer,1, 2, a) J.-F. Boussuge,1, b) P. Sagaut,3, c) M. Montagnac,1 D.3
Papadogiannis,2 and X. Garnaud44
1)CERFACS, 42 Avenue Gaspard Coriolis, 31057 Toulouse,5
France6
2)SAFRAN Tech, Rue des Jeunes Bois, Chateaufort – CS 80112,7
78772 Magny-les-Hameaux, France8
3)Aix Marseille Univ, CNRS, Centrale Marseille, M2P2 UMR 7340,9
13451 Marseille, France10
4)SAFRAN Aircraft Engines, Rond-point Rene Ravaud, 77550 Moissy-Cramayel,11
France12
(Dated: 25 May 2018)13
1
Wall-modeling is required to make large-eddy simulations of high-Reynolds num-
ber wall-bounded turbulent flows feasible in terms of computational cost. Here, an
extension of the integral wall-model for large-eddy simulations (iWMLES) for in-
compressible flows developed by Yang et al. [Phys. Fluids 27, 025112 (2015)] to
compressible and isothermal flows is proposed and assessed. iWMLES approach is
analogous to the Von Karman-Pohlhausen integral method for laminar flows: the ve-
locity profile is parametrized and unknown coefficients are determined by matching
boundary conditions obeying the integral boundary layer momentum equation. It
allows to include non-equilibrium effects such as pressure gradient and convection at
a computing cost similar to analytical wall-models. To take into account density vari-
ations and temperature gradients, the temperature profile is also parametrized and
the integral compressible boundary layer energy equation is considered. Parametrized
profiles are based on the usual logarithmic wall functions with corrective terms to
extend their range of validity. Instead of solving a set of differential equations as wall-
models based on the thin boundary layer equations approach, a simple linear system
is solved. The proposed wall-model is implemented in a finite-volume cell-centered
structured grid solver and assessed on adiabatic and isothermal plane channel flows
at several friction Reynolds and Mach numbers. For low Mach number cases, mean
profiles, wall fluxes and turbulent fluctuations are in agreement with those of DNS.
For supersonic flows, the results are in good agreement with the DNS data, espe-
cially the mean velocity quantities and the wall friction while standard analytical
wall-models show their limits.
a)Electronic mail: [email protected])Electronic mail: [email protected])Electronic mail: [email protected]
2
I. INTRODUCTION14
Computational fluid dynamics is widely used as a research or engineering tool in fluid15
mechanics. Reynolds-Averaged Navier-Stokes (RANS), which is based on statistical tur-16
bulence models, is the most popular paradigm. RANS modeling allows to compute mean17
quantities at a reasonable accuracy and computational cost even for complex engineering18
flows1. But these models show their limits to predict highly unsteady phenomena like sep-19
aration. Large-Eddy Simulation (LES) is a way to overcome these limitations and the next20
step toward achieving more accurate simulations2–4. In RANS simulations the whole tur-21
bulent spectrum is modeled while LES captures the large energetic scales of the flow and22
only the small scales are modeled. However, near-wall turbulence contains very small scales23
(near-wall streaks5) and simulating them would require prohibitively fine meshes. A first24
estimation of LES computational cost, with or without wall-modeling, was given by Chap-25
man 6 . Using a more accurate correlation for high Reynolds number flows, Choi and Moin 726
recently updated these estimations: the number of mesh points to solve a boundary layer27
flow is approximatively proportional to Re13/7 in the inner layer and Re1 in the outer layer28
with Re the Reynolds number. In typical aeronautical applications, one has Re ∼ 107, thus29
the inner layer requires more than 99% of the boundary layer grid points as illustrated by30
Piomelli and Balaras 8 .31
Hybrid RANS/LES methods are a way to reduce this computational cost. Several reviews32
of these methods exist8–11. The two most popular are the Detached-Eddy Simulation (DES12)33
and the Wall-Modeled LES (WMLES2,13,14). Both methods have pros and cons but the same34
goal: to tackle LES limitations to high Reynolds number flows by using RANS modeling35
near walls. The number of required grid points, N , was estimated15,16 to N ∼ Re2.17θ and36
N ∼ Re1.17θ for wall-resolved and wall-modeled LES of flat plate boundary layer.37
On the one hand, DES takes advantage of the similarity of the LES and RANS equations:38
in the near wall region the RANS equations are solved and when the mesh is fine enough, the39
turbulent viscosity switchs to a subgrid-scale model and LES behaviour is obtained. DES40
was historically developed with the Spalart-Allmaras turbulence model17 but other models41
are now available18. However, the transition region between RANS and LES modeling can be42
a cause of errors. Indeed non-physical terms are generated near the transition region19. To43
cure this issue and depending on the case, stochastic forcing20,21 methods can be calibrated44
3
and used in the transition region to cancel these terms and accelerate the transition between45
RANS and LES.46
On the other hand, WMLES solves LES equations in all the computational domain but as47
opposed to a resolved LES, the mesh is coarsened in the vicinity of walls. Then, a wall-model48
is used to compute the wall fluxes to take into account the effects of the inner layer part of49
boundary layers. Wall-models use LES data near walls as input. So the near-wall treatment50
in WMLES is very localized which eases implementation, efficient parallel computation and51
application to complex geometries. However, due to a coarse mesh near walls, numerical52
methods and subgrid-scale model must be carefully chosen22,23 to avoid the development53
of non-physical turbulent structures. Indeed, by definition, LES2,3 solves the largest eddies54
and models the smallest ones thanks to a subgrid-scale model. The frequency cutoff is55
generally given by the cell size of the computational grid and should be located within the56
inertial range of the turbulent energy spectrum. Due to this restriction, near-walls cells57
in resolved LES must be fine, as the turbulent structures inside boundary layers are very58
small5,24. Eddies not captured by the grid are then taken into account by the subgrid-scale59
model, often assuming they are isotropic. In WMLES the local LES filter is not expected60
to fall in the inertial range because of the coarse near-walls cells, thus, there is no reason61
to accurately predict the subgrid-scale viscosity. Corrections of the near-wall subgrid-scale62
viscosity have been proposed by several authors23,25,26. Bocquet, Sagaut, and Jouhaud 2363
show in a WMLES of a plane channel flow that without correction the subgrid-scale term64
is highly overestimated in the first off-wall cell.65
Another consequence of the use of a coarse grid near the wall in WMLES is the potential66
under-resolution of the turbulent streaky structures5 located in the logarithmic region. In-67
deed in WMLES, inner layer effects are modeled by the wall-model and only the outer layer is68
directly solved. Therefore, streaks in the viscous sublayer are not captured. The overlap (log-69
arithmic) region is mainly composed of large-scale structures: Large-Scale Motions (LSM27)70
which have streamwise sizes of lx ∼ 2–3δ and the Very-Large-Scale Motions (VLSM28,29)71
whose sizes scale with lx ∼ 10δ with δ the boundary layer thickness. Since, in the outer72
region, a major contribution of the turbulent kinetic energy and Reynolds stress is due to73
these large structures, they must be well resolved. The number of grid points per wavelength74
required to resolve these structures depends on the numerical methods employed22. In the75
wall-normal direction, the wall-model input data should be placed in the overlap region76
4
i.e. at y ' 0.05–0.2δ. If the turbulent motions are under-resolved, non-physical turbulent77
structures are present near walls like the super-streaks encountered in DES or other hybrid78
RANS/LES methods13,20. However, if the grid is too fine, under-resolved near-wall streaks79
start to be captured and are used as input to steady wall-model equations8. Therefore, the80
physical and numerical modeling of the flow must be carefully chosen to accurately compute81
the skin friction. Otherwise the well-known log-layer mismatch appears13.82
In the following, only WMLES is considered. Two types of wall-models exist: analytical83
and numerical. On the one hand, analytical wall-models are the simplest ones. Both the84
velocity and the temperature at a given point (usually the first off-wall point) are assumed85
to follow a mean solution. The treatment’s cost is negligible compared to a LES time step86
but the range of validity of an analytical wall-model is restricted to simple flows. Note87
that analytical wall-models using the standard logarithmic law of the wall10 are also called88
equilibrium wall-models since this law is derived assuming a constant shear stress in the89
inner layer part of the turbulent boundary layer. Moreover, the logarithmic law of the wall90
is in theory only valid for a steady incompressible flow along a fully turbulent boundary91
layer at zero pressure gradient.92
On the other hand, numerical wall-models solve a set of differential equations usually93
based on the thin boundary layer equations approximation under the RANS formalism on94
an implicit or separated30 grid between the wall and the wall-model interface. Therefore, the95
effects of the turbulent structures located in the inner layer are taken into account globally8.96
Numerical wall-models have a larger domain of validity than analytical wall-models since97
they can take into account more physical terms: advection, unsteadiness, pressure gradient,98
etc. The cheapest and simplest numerical wall-models (called equilibrium wall-models)99
only consider the wall-normal diffusion. Unsteadiness are assumed to be negligible and100
pressure gradient and convection terms balance exactly. Nevertheless, this hypothesis can101
lead to inaccurate estimation of the wall fluxes in non-equilibrium flows. To improve the102
wall fluxes prediction, non-equilibrium wall-models are thus required. To go beyond the103
equilibrium assumption, non-equilibrium wall-models take also into account the pressure104
gradient23 and possibly the time derivative31. Bidimensional convective terms can also be105
added32. Most complete numerical wall-models consider the full three-dimensional unsteady106
RANS thin boundary layer equations25,33–35. Indeed, it is important to keep all terms of the107
thin boundary layer equations, since pressure gradient and convective terms are coupled36.108
5
However, the computational cost of numerical wall-models is also higher due to the numerical109
resolution of complex two or three-dimensional equations. Besides, to add non-equilibrium110
terms such as wall-parallel diffusion or convective terms, each wall-adjacent cell should111
exchange data with its neighbors, reducing the parallel efficiency of the CFD solver especially112
for unstructured solvers30.113
To take into account compressibility effects and/or temperature gradient, the current114
state-of-the-art in WMLES is to use an equilibrium numerical wall-model. Indeed, such115
kind of wall-model has been sucessfully used on:116
• incompressible and compressible plane channel flows with or without hear transfer23 ;117
• supersonic boundary layers flows22 ;118
• oblique shock/boundary-layer interaction37 ;119
• supersonic Couette flows with heat transfer38.120
However, in flows with many walls (e.g. in turbomachinery flows), the computational over-121
head of such wall-models can be important. This is especially the case at very high Reynolds122
number, since the computational cost of numerical wall-models is often non-linear with re-123
spect to the Reynolds number due to the increased number of degrees of freedoms required124
to their resolution.125
In the frame of this work, the aim is to develop a wall-model capable of taking into126
account the physical phenomena encountered in turbomachinery systems. Usually, in tur-127
bomachinery flows, Mach numbers are up to 1.539–43, Reynolds numbers44,45 between 105128
and 107 and temperature gradients in the order of ten to hundred Kelvin46. As pointed out129
by Tyacke and Tucker 47 , if the flow in low-pressure turbines can be simulated with LES130
due to their relatively low-Reynolds number, the flows in the other components require a131
wall-modeling in order to reduce computational requirements at high Reynolds number. The132
interest of a wall-modeling for compressor flows is also highlighted by Gourdain et al. 48 .133
Recently, an hybrid wall-model, namely integral Wall-Model for Large-Eddy Simulation134
(iWMLES), has been developed for adiabatic incompressible flows49. In iWMLES the lon-135
gitudinal velocity profile is parametrized and imposed to satisfy some boundary conditions136
and among them the vertically integrated momentum thin boundary layer equation. This137
6
method allows to include more physical phenomena than an analytical wall-model, like pres-138
sure gradient, roughness effects or curvature effects while keeping the computational cost139
very low compared to standard numerical wall-model as only a weak solution is sought.140
iWMLES has been applied to attached boundary layer flows with a roughness surface49 or141
separated flow along a smooth surface50.142
In this study, an extension of iWMLES for compressible and isothermal flows on smooth143
walls is proposed. In Section II the flow solver and the wall-model implementation are144
described. Then, in Section III, the iWMLES approach is explained along with its numerical145
resolution. Finally, in Section IV the integral wall-model developed in this study, is tested on146
quasi-incompressible and supersonic bi-periodic plane channel flows, whose Mach numbers147
and temperature gradients are representative of typical turbomachinery flows. Wall fluxes,148
velocity, temperature mean profiles and fluctuations are compared to DNS data.149
II. DESCRIPTION OF THE FLOW SOLVER150
Even though wall-models can be seen as black boxes which compute wall fluxes using151
LES data such as the velocity, pressure or temperature at some distance from the wall, they152
are strongly coupled with the flow solver. Indeed numerical choices must be made on how:153
(i) the LES data are used as inputs of the wall-model ; (ii) the outputs of the wall-model154
are used by the LES flow solver ; (iii) to take into account the coarse near walls cells.155
A. Physical and numerical modeling of the flow156
The CFD solver used for this study (elsA-ONERA51) is based on a cell-centered finite-157
volume approach to solve LES equations. The fluid is assumed to be a perfect gas and the158
viscosity µ varies with the temperature T according to the Sutherland’s law:159
µ(T ) = µref
(T
Tref
)3/2Tref + S
T + S(1)
with Sutherland’s constant S = 110.4 K. µref and Tref are respectively a reference viscosity160
and a reference temperature.161
Compressible LES equations are written under the Favre averaging and on implicit mesh162
filtering formalism. In the following, RANS and LES variables would be denoted respectively163
with a bar (f) and a tilde (f).164
7
The subgrid scale-model is estimated by the Wall-Adapting Local Eddy-viscosity (WALE52)165
subgrid-scale model which is based on the velocity gradient:166
νsgs = (Cw∆)2(SdijSdij)3/2
(SijSij)5/2 + (SdijSdij)5/4(2)
with the constant Cw = 0.5, ∆ the cell volume, ui the velocity components, Sij = (∂ui/∂xj+167
∂uj/∂xi)/2 the components of the strain rate tensor and168
Sdij =1
2
(∂ui∂xl
∂ul∂xj
+∂uj∂xl
∂ul∂xi
)+
1
3
∂um∂xl
∂ul∂xm
δij (3)
where δij is the Kronecker symbol.169
In the finite volume framework, thanks to the Gauss divergence theorem, the divergence170
of the convective and diffusive terms of the Navier-Stokes equations is reduced to a sum of171
fluxes. Convective fluxes are estimated using a second-order centered scheme without arti-172
ficial dissipation in the skew-symmetric form, which is observed to reduce aliasing errors53.173
To determine the diffusive fluxes, the laminar viscosity, the subgrid-scale model, the velocity174
gradient and the temperature gradient at each interface between cells are required. Laminar175
viscosity and subgrid-scale model at interface are evaluated by averaging the two adjacent176
cells values. Velocity and temperature gradients are computed using the Gauss divergence177
theorem on a shifted control volume (centered on the interface). This discretization leads178
to a three point stencil second-order scheme. At walls, two fictitious cells are introduced in179
order to keep the same schemes.180
When dealing with moderate or high Reynolds number flows, the convective fluxes are181
expected to be dominant over the diffusive fluxes, except in the walls vicinities. However,182
in WMLES, the velocity and temperature gradients discretization are erroneous at walls183
interfaces since the near-wall coarse mesh used cannot capture high gradients. Therefore,184
velocity and temperature gradients at each wall interface are evaluated using a wall-model185
as explained in the next subsection II B.186
Time advancement is performed explicitely using a low storage Runge-Kutta scheme with187
8
four stages. For a differential equation ∂u/∂t = f(u, t), the steps are given by:188
u(0) = un
u(1) = un + ∆t4R(0)
u(2) = un + ∆t3R(1)
u(3) = un + ∆t2R(2)
un+1 = un +R(3)
(4)
with ∆t the time step and R(i) the intermediate residuals defined by:189
R(0) = f(u(0), tn)
R(1) = f(u(1), tn + ∆t4
)
R(2) = f(u(2), tn + ∆t3
)
R(3) = f(u(3), tn + ∆t2
)
(5)
This scheme is fourth-order if f is linear, second-order otherwise. In all the following simula-190
tions, a constant time step is chosen such that the Courant-Friedrichs-Lewy (CFL) number191
is around 0.7.192
This numerical framework has already been proposed and validated23, except here the193
WALE subgrid-scale model is employed instead of the selective Smagorinsky model to better194
predict supersonic flows.195
B. Relation between the flow solver and the wall-model196
Due to the coarse mesh used in WMLES, high velocity and temperature gradients at197
walls cannot be properly captured. That is why at each time step, during the viscous flux198
computation, the wall fluxes are computed by the wall-model. Given information from the199
LES field at the matching point (the first off-wall cell in the frame of this work) like the200
velocity vector, pressure, density or temperature, the wall friction vector τw and the wall201
heat flux φw are returned by the wall-model. Note that τw and φw are not exactly LES202
variables as they are the output of RANS-based wall-model equations.203
The wall friction vector τw is assumed to be aligned with the velocity vector at the first204
off-wall point projected in the plane parallel to the wall: u1‖ . Thereby, only two-dimensional205
9
equations can be considered in wall-models, reducing their complexity and computational206
cost. The projection step is briefly reminded here. At each call of the wall-model, the207
velocity vector at the first off-wall point u1 is projected in the plane parallel to the wall as208
shown in Fig. 1, with n the wall-normal vector, y1 the distance between the wall and the209
first off-wall cell and t = u1‖/∥∥u1‖
∥∥ the local velocity direction. The wall-model equations210
are solved in the local boundary layer frame (t,n) and the wall friction modulus ‖τw‖ and211
the wall heat flux are returned to the LES solver. As τw is assumed to be aligned with u1‖ ,212
the wall friction vector coordinates in the flow solver frame can be computed and used as a213
boundary condition by the flow solver. It can be noticed that from the wall-model point of214
view, the boundary layer is always attached.215
t ∧ nt
n
(u1 · n)nu1
y1u1‖
τw
‖τw‖ /µw
ji
FIG. 1: Scheme of the projection of the first off-wall point velocity onto the local boundary
layer frame
C. Analytical wall-model216
A fair comparison between wall-models can only be done in a given flow solver, since217
there is a strong coupling between the two. Therefore, a standard analytical wall-model is218
implemented in order to assess the wider range of validity of numerical wall-models with219
respect to analytical ones. It is based on the turbulent boundary layer equations under zero220
pressure gradient for an incompressible flow. Turbulence is assumed to be statistically bidi-221
mensional and the mean flow steady. Under these hypotheses, the wall-normal momentum222
equation implies that the wall-normal variation of pressure is of order of the boundary layer223
10
thickness. So this variation can be neglected with respect to the streamwise pressure gradi-224
ent and the pressure is assumed to be constant along the wall-normal direction. Moreover,225
by an order-of-magnitude analysis it can be shown that in the limit of infinite Reynolds226
number, the shear stress τ is constant in the inner layer part of the boundary layer and227
therefore equal to the wall shear stress τw. This result is the key point to justify the use228
of wall-laws because a single velocity scale uτ =√‖τw‖ /ρw in all the inner layer can be229
defined and a velocity profile u+ = u/uτ , independant of the outer layer, can be sought.230
With this velocity scale, a linear profile u+ = y+ in the linear sublayer and a logarithmic231
profile u+ = (1/κ) ln y+ + B in the inertial sublayer are found. Similarly, for an isother-232
mal wall, the heat flux φ is constant and equal to the wall heat flux φw in the inner layer233
region and a temperature scale Tτ = φw/(ρwcpuτ ) can be defined with cp the specific heat234
capacity at constant pressure and ρw the wall-density. A self-similar temperature profile235
T+ = (T − Tw)/Tτ following a linear law in the linear sublayer and a logarithmic law in the236
inertial sublayer is found, where Tw is the wall temperature.237
Based on these results, Reichardt 54 and Kader 55 proposed smooth velocity and tem-
perature wall-laws in the whole inner layer. They provide an explicit relation between the
mean velocity and temperature and the wall fluxes ‖τw‖ and φw. From these two laws, an
analytical wall-model23,56 can be built:
u+1‖
= u+1 =
u1
uτ=
1
κln(1 + κy+
1 ) + (B − 1
κlnκ)(1− exp(
−y+1
11)− y+
1
11exp(−y+
1
3)) (6)
T+1 = T+
1 =T 1 − TwTτ
= Pry+1 exp(−Γ) + (
1
κtln(1 + y+
1 ) + β) exp(−1/Γ) (7)
with Pr the Prandtl number and
κ = 0.41 (8a)
B = 5.25 (8b)
κt =1
2.12(8c)
Γ =10−2(Pry+
1 )4
1 + 5Pr3y+1
(8d)
β = (3.85Pr1/3 − 1.3)2 +1
κtln(Pr) (8e)
238
As explained in subsection II B, the bidimensional unsteady velocity u1‖ and temperature239
T1 at the first off-wall cell are imposed to follow a mean wall-law at each time step. Given240
11
LES data at the first off-wall point, the following procedure is applied in order to compute241
the wall fluxes:242
1. The pressure is assumed to be constant in the wall-normal direction, allowing to de-243
termine the wall density thanks to the perfect gas relation ρw = p1/(rTw) with p1 the244
pressure at the first point, and r the perfect gas constant. In the case of an adiabatic245
boundary condition the wall temperature is supposed equal to the friction temperature246
Tτ = Te(1 + Pr1/3(γ − 1)/2Me‖
2). As the values at the outer edge of the boundary247
layer like temperature Te and Mach number Me‖ = ue‖/
√γrTe are not accessible in248
complex geometries, their value at the first off-wall cell are taken.249
2. Before being used as input by the wall-model, the velocity vector is projected along250
the plane parallel to the wall as explained in subsection II B.251
3. The friction velocity uτ =√‖τw‖ /ρw is computed thanks to Eq. (6) and a Newton-252
Raphson algorithm.253
4. For an isothermal wall, the friction temperature Tτ = −φw/(ρwcpuτ ) is computed254
thanks to Eq. (7) and a Newton-Raphson algorithm.255
5. The wall friction vector components in the flow solver frame are computed by switching256
from the boundary layer frame to the flow solver frame.257
In the following, this analytical wall-model will be noted aWMLES.258
III. COMPRESSIBLE INTEGRAL WMLES APPROACH259
In this section, an integral wall-model for LES (iWMLES) of compressible flows along260
adiabatic or isothermal smooth plates is presented. First, the key idea of the wall-model is261
described. As it is based on the integral thin boundary layer equations, these equations are262
formulated before explaining how to solve them. Finally, the overall wall-model resolution263
is summarized.264
12
A. Description of the iWMLES approach265
For a laminar boundary layer, von Karman57 and Pohlhausen58 proposed a method to266
estimate the local velocity profile without solving differential equations. Instead of directly267
solving the boundary layer equations, the longitudinal velocity profile is approximated by268
a fourth-order polynomial. Then, the only unknowns to fully determine the velocity profile269
are the polynomial coefficients. They are computed to impose the local longitudinal velocity270
to verify physical boundary conditions and among them the vertically integrated boundary271
layer momentum equations. Thus, a velocity profile taking into account pressure gradient272
is built at a very low computational cost. In the case of laminar boundary layer with zero273
pressure gradient, the friction coefficient computed presents 2% of error compare to the274
exact Blasius solution.275
Here, a formulation of iWMLES for compressible and isothermal flows on smooth walls is276
proposed and assessed. This wall-model is analogue to the Von Karman-Pohlhausen integral277
method except the boundary layer is assumed to be fully turbulent. To handle these kind278
of flows, the velocity and the temperature profiles are parametrized. Then, the unknown279
coefficients are computed so that velocity and temperature profiles satisfy some boundary280
conditions and among them the vertically integrated bidimensional turbulent boundary layer281
momentum and energy equations. Therefore the computed profiles represent a weak solution282
of the integral boundary layer equations and take into account more physics than analytical283
wall-models. Instead of solving directly the thin boundary layer equations and use expensive284
numerical algorithms, a weak solution is sought and a simple scalar equations system has to285
be solved. So iWMLES can be seen as a hybrid wall-model: either an analytical wall-model286
with dynamic coefficients or a numerical wall-model where only a weak solution is searched.287
B. Integral compressible thin boundary layer equations288
In this subsection the vertically integrated boundary layer momentum and energy equa-
tions required in iWMLES are formulated. In the following the longitudinal and normal
velocities are respectively u and v, the longitudinal and wall-normal directions are respec-
tively the x and y axis. The other variables follow conventional nomenclature. As opposed to
the original formulation of iWMLES for incompressible flows49, only bidimensional equations
13
are considered here due to the projection explained in subsection II B. The two-dimensional
turbulent thin boundary layer equations in cartesian coordinates for a perfect gas are:
∂ρ
∂t+∂ρu
∂x+∂ρv
∂y= 0 (9a)
∂ρu
∂t+∂ρuu
∂x+∂ρvu
∂y= −∂p
∂x+∂τ
∂y(9b)
∂p
∂y+∂ρv′2
∂y= 0 (9c)
∂ρcpT
∂t+∂ρucpT
∂x+∂ρvcpT
∂y=∂p
∂t+ u
∂p
∂x+ τ
∂u
∂y− ∂φ
∂y(9d)
p = ρrT (9e)
with the shear stress τ = (µ + µt)∂u∂y
and the heat flux φ = −(λ + λt)∂T∂y
, λ = cpµ
Pris289
the molecular thermal conductivity. µt and λt are the turbulent viscosity and the thermal290
conductivity respectively, v′ is the fluctuating part of the wall-normal velocity. Equation (9a)291
is related to the conservation of mass, Eqs. (9b) and (9c) are the longitudinal and wall-normal292
momentum conservation equations. Equation (9d) is the energy equation and the perfect293
gas law (9e) closes the equations system. In the above equations the diffusion term along294
the longitudinal direction x is neglected.295
In the following, for the sake of simplicity, any vertically integrated quantities f between296
y = 0 and y = y1 will be denoted297
Lf =
∫ y1
0
f dy
According to Eq. (9c), the wall-normal pressure gradient is of order of magnitude of298
the boundary layer thickness and therefore is small compared to the streamwise pressure299
gradient. This involves an important result in the boundary layer theory: the pressure can300
be assumed to be constant along the wall-normal direction, which means it is an input of the301
wall-model extracted from the LES field. Taking advantage of this approximation and the302
perfect gas law (9e), Eqs. (9a), (9b) and (9d) are integrated along the wall-normal direction303
between the wall (assumed without loss of generality to be located at y = 0) and the first304
off-wall point at y = y1:305
14
ρ1v1 = −∫ y1
0
(∂ρ
∂t+∂ρu
∂x
)dy (10a)
∂Lρu∂t− u1‖
∂Lρ∂t
+Mx = τ1 − τw (10b)
(cpr− 1) ∂p1
∂ty1 − cpT1
∂Lρ∂t
+MTx = Lτ ∂u∂y− φ1 + φw (10c)
with u1‖ , T1 and p1 the LES data at y1. The convective terms are expressed as follows:
Mx =∂p1
∂xy1 +
∂Lρu2
∂x− u1‖
∂Lρu∂x
(11)
MTx =cprp1∂Lu∂x− cpT1
∂Lρu∂x− ∂p1
∂xLu (12)
Scalar equations (10b) and (10c) are respectively the vertically integrated momentum and
energy equations. Velocity and temperature profiles in iWMLES will be constrained to be
solutions of these equations. Note that these equations require only the values of turbulence
models at the first off-wall cell and of the vertically integrated shear stress. Thus, iWMLES
is less sensitive to turbulence models than in numerical wall-models where turbulence models
are involved over the entire wall-normal direction and need to be carefully chosen23,25,33,34.
Therefore, standard mixing-length models are used:
µt(y) = ρ(κyVD26(y))2|∂u∂y|(y) (13)
λt(y) = ρcp(κy)2VD26(y)VD35(y)|∂u∂y|(y) (14)
with VD the Van-Driest damping function defined by:
VDa(y) = 1− exp
(−y+3
a
), ∀a ∈ R∗ (15)
y+3 (y) =
yρ√τw/ρ
µ=yuτ√ρwρ
µ(16)
An y+3 scaling (also denoted y∗ in the literature) is used to improve compressibility effects306
prediction in Eqs. (13) and (14). Using this scaling, a better prediction of the wall fluxes on307
supersonic boundary layers is obtained59 either for adiabatic or isothermal walls. A better308
behaviour is also observed for velocity fluctuations and Reynolds stress60,61 with the y+3309
scaling compare to y+.310
15
C. Parametrized profiles311
Instead of directly solving Eqs. (10b) and (10c), a weak solution is sought by parametriz-312
ing velocity and temperature profiles. Analytical results show that for a steady incom-313
pressible fully turbulent boundary layer flow with zero pressure gradient, the longitudinal314
velocity follows a logarithmic law. Even if the existence and the universality of this law315
are still debated62,63, it is observed to hold even for low Mach number flow with or without316
pressure gradient64,65 or curvature effects66.317
Thus, the velocity and temperature profiles are parametrized based on the logarithmic318
laws of the wall defined in Eqs. (6) and (7) with corrective terms:319
u(y;uτ ; δν ;A) = uτ
(1
κln(1 + κy+) + (B − 1
κlnκ)(1− exp(
−y+
11)− y+
11exp(−y+
3))
+ Ay+(1− exp(−y+
δν))
) (17)
T (y;uτ ;Tτ ; δc;AT ) = Tw + Tτ
(Pry+ exp(−Γ) + (
1
κtln(1 + y+) + β) exp(−1/Γ)
+ ATy+(1− exp(
−y+
δc))
)for isothermal wall
T (y;u) = T1
(1 + Pr 1/3 γ−1
2M2
1‖(1− u2
u21‖)
)for adiabatic wall
(18)
with M1‖ = u1‖/
√γrT1 the local Mach number based on the velocity projected in the plane320
parallel to the wall (u1‖), A, AT , δc, δν , uτ and Tτ being six scalars unknown.321
The correction terms A andAT aim to simulate the deviation from the logarithmic law and322
build a local dynamic wall-law which satisfies the vertically integrated momentum and energy323
equations (10b) and (10c). A linear correction term is added to the temperature profile and324
a square root correction term is added to the velocity profile as previous studies have shown325
that a square root region appears when the boundary layer is not at equilibrium67,68. Since326
the viscous sublayer seems to be only slightly affected by non-equilibrium effects, correction327
terms are applied only in the inertial sublayer thanks to a Van-Driest damping function328
1− exp(−y+/δν).329
For an adiabatic boundary condition the wall heat flux is zero and only the wall friction330
vector needs to be determined by the wall-model. Moreover, the temperature profile does not331
follow the logarithmic law of the wall. Thus, in this case, the temperature is determined from332
16
Walz’s law69. In theory, this relation needs the values at the outer edge of the boundary layer,333
but these quantities are not easily accessible in complex geometries or in unstructured solvers.334
Local values at first off-wall cell are then used, and this approximation will be validated a335
posteriori. Besides, as the wall heat flux does not need to be computed, the vertically336
integrated energy equation (10c) is not solved as it will be explained in subsection III D.337
As the y+ scaling does not take into account the wake region, selected profiles and tur-338
bulence models (13) and (14) are only valid in the inner layer of the boundary layer. From339
a physical viewpoint, improved stability and accuracy are obtained when the wall-model340
interface is located in the overlap region between the inner and outer layers8,13. Thus, the341
wall-model interface is aimed to be about at 5–15% of the boundary layer thickness and342
y+ & 50. This is also the case of the analytical wall-model defined by Eqs. (6) and (7).343
D. Evaluation of the profile parameters: selected boundary conditions344
To completely define the velocity and temperature profiles given by Eqs. (17) and (18),345
three scalars unknowns must be computed in the case of an adiabatic wall: uτ , δν and A. If346
the wall is considered to be isothermal, there are three additional scalars unknowns: Tτ , δc347
and AT . They are determined according to boundary conditions and vertically integrated348
boundary layer equations:349
1. the velocity profile defined in Eq. (17) must match with the LES velocity field at y1350
in order to be continuous:351
u1(y1;uτ ; δν ;A) = u1‖ (19)
2. the diffusive layer thickness δν is supposed to be constant:352
δν = 11 (20)
This value corresponds to the intersection point between the linear law u+ = y+ and353
the logarithmic law of the wall in the case of an incompressible flow ;354
3. the velocity profile defined in Eq. (17) must be solution of the vertically integrated355
momentum equation (10b) ;356
4. in the case of an isothermal boundary condition, three additional unknowns AT , δc357
and Tτ must be computed:358
17
(a) the temperature profile defined in Eq. (18) must match with the LES temperature359
field at y1 in order to be continuous:360
T 1(y1;uτ ;Tτ ; δc;AT ) = T1 (21)
(b) as with the diffusive layer thickness δν , the convective layer thickness δc is sup-361
posed to be constant and362
δc = 11 (22)
(c) the velocity and temperature profiles defined in Eqs. (17) and (18) must be solu-363
tion of the vertically integrated energy equation (10c).364
All these equations form a system of three or six (depending on whether the wall is365
supposed to be adiabatic or isothermal) scalars equations:366
uτ
(1
κln(1 + κy+
1 ) + (B − 1
κlnκ)(1− exp(
−y+1
11)− y+
1
11exp(−y+
1
3))
)
+Ay+1 (1− exp(
−y+1
δν)) = u1‖
δν = 11
∂Lρu∂t− u1
∂Lρ∂t
+Mx = τ1 − τw
(23a)
and if the boundary condition is isothermal:367
Tw + Tτ
(Pry+
1 exp(−Γ) + (1
κtln(1 + y+
1 ) + β) exp(−1/Γ)
+ATy+1 (1− exp(
−y+1
δc))
)= T1
δc = 11(cpr− 1) ∂p1
∂ty1 − cpT1
∂Lρ∂t
+MTx = Lτ ∂u∂y− φ1 + φw
(23b)
which is much easier to solve than numerical wall-models differential equations.368
For an adiabatic boundary condition, the list of unknowns is reduced to three as only the369
wall friction vector has to be computed by the wall-model. Indeed in this case the friction370
temperature Tτ is zero as well as the wall heat flux. As explained in subsection III C, Walz’s371
law69 is used to link the temperature to the velocity profile and the correction term AT is372
not required. Therefore, in this case, only the velocity profile needs to be determined with373
the friction velocity uτ , the diffusive layer thickness δν and the correction term A.374
18
E. Numerical resolution of the integral equations375
The method to solve the system of equations (23) written in subsection III D is now376
detailed in the case of an isothermal wall. For an adiabatic wall, the resolution is simpler377
as the temperature profile is determined from the velocity profile using Eq. (18). In this378
section, variables at the current time step will be denoted with a superscript n and n − 1379
for those at previous time step.380
At each wall interface, using LES data located at the first off-wall cell y1 and thanks to381
Eqs. (19) to (22), the scalars unknowns An, δν , AnT and δc are expressed as functions of unτ382
and T nτ :383
An(unτ ;T nτ ) =
un1‖unτ−(
1κ
ln(1 + κy+1n) + (B − 1
κlnκ)(1− exp(
−y+1n
11)− y+1
n
11exp(
−y+1n
3)))
y+1n(1− exp(
−y+1n
δν))
δν = 11
AnT (unτ ;T nτ ) =
Tn1 −TwTnτ−(
Pry+1n
exp(−Γ) + ( 1κt
ln(1 + y+1n) + β) exp(−1/Γ)
)
y+1n(1− exp(
−y+1n
δc))
δc = 11
(24)
Thus, the system (23) is reduced to a non-linear system of two equations ((10b) and
(10c)) with two unknowns: unτ and T nτ . The temporal derivatives in these equations are
discretized with an explicit scheme and the spatial derivatives with a second-order centered
scheme. An upwind scheme is used if the wall interface is adjacent to another boundary
condition. Then Eqs. (10b) and (10c) become:
Lnρu(unτ ;T nτ )− un−1
1‖Lnρ(unτ ;T nτ ) = Ln−1
ρu (1− un−11‖
) + ∆t(−Mn−1x + τ1
n−1 − τwn−1)
cpTn−11 Lnρ(unτ ;T nτ ) =
(cpr− 1)y1(pn1 − pn−1
1 ) + cpTn−11 Ln−1
ρ
−∆t(−Mn−1Tx
+ Ln−1
τ ∂u∂y
− φ1n−1 + φw
n−1)
(25a)
(25b)
where the terms depending on the unknowns unτ and T nτ are specified and with ∆t the384
time step of the LES solver. Given data located at the first off-wall point y1 at current and385
previous time step along with the wall fluxes at previous iteration un−1τ and T n−1
τ , unτ and T nτ386
are computed by solving Eqs. (25a) and (25b) thanks to a bidimensional Newton-Raphson’s387
algorithm whose the corresponding Jacobian matrix is given in Appendix A. un−1τ and T n−1
τ388
19
are chosen as initial guesses of the algorithm. At the first iteration of the flow solver, i.e. at389
n = 0, wall fluxes are estimated by a first-order finite difference scheme like in wall-resolved390
LES.391
Integral terms Lf in Eqs. (25a) and (25b) cannot be determined analytically unless the392
density is constant49. Besides, since the pressure is assumed to be constant along the wall-393
normal direction, density profile can be determined from the temperature profile using the394
perfect gas law. Thus, a Gauss-Legendre quadrature is used to compute the Lf terms:395
definite integrals are approximated as a weighted sum of functions values at N points. For396
a function f : [0 ; y1]→ R, the quadrature implies397
∫ y1
0
f(y) dy ' y1
2
N∑
i=0
wif(y1
2(1 + ξi)
)(26)
with (wi)0≤i≤N and (ξi)0≤i≤N respectively the weights and the Gauss nodes. In the case of398
the Gauss-Legendre quadrature, wi and ξi are chosen in such a way that the quadrature is399
exact for Legendre polynomials of degree less or equal to 2N − 1.400
Figure 2 shows the observed computational overhead of iWMLES and an equilibrium401
numerical wall-model23 (denoted TBLE in the following) with respect to an analytical model402
aWMLES in the cases of an isothermal or adiabatic wall. The iWMLES computational403
cost is much lower compared to the TBLE one, although the former takes into account404
convective terms. Besides, and as opposed to TBLE, iWMLES algorithm scales linearly405
with the number of Gauss nodes N used. The iWMLES algorithm is nearly twice as fast406
when the wall is assumed to be adiabatic instead of isothermal. Indeed, in this case the407
integral energy equation is not solved and a velocity-temperature relation is used. Such408
improvement is also observed for TBLE but in a reduced proportion as the temperature409
profile still needs to be computed. In the following N is taken equal to 10.410
The iWMLES resolution algorithm is illustrated in Fig. 3 and can be summarized as411
follows:412
1. At the first iteration, wall fluxes are computed by a nearest-cell approximation as in413
standard resolved LES.414
2. Then, at each wall interface the LES data at matching point y1 at previous and current415
iteration n along with the wall fluxes at previous iteration n− 1 are used as input to416
the iWMLES model in order to compute the wall fluxes at the current iteration:417
20
100
101
102
103
104
105
106
0 20 40 60 80 100
Twall-m
odel
TaW
MLES
Number of Gauss points
iWMLES - adiabatic wall
iWMLES - isothermal wall
TBLE - adiabatic wall
TBLE - isothermal wall
FIG. 2: Ratio of the observed computational cost Twall-model of the iWMLES and a
numerical wall-model for LES (TBLE) based on the equilibrium thin-boundary layer
equations with respect to the cost TaWMLES of an analytical wall-model per cell interface
for adiabatic and isothermal walls. For iWMLES, corresponding linear regression are in
dashed lines.
(a) As with the analytical wall-model described in subsection II C, wall density ρnw is418
computed from the perfect gas relation.419
(b) The velocity at the current iteration n is projected along the plane parallel to420
the wall as explained in subsection II B.421
(c) At each wall interface and given data at the previous solver iteration, the scalars422
parameters An−1, An−1T , δν and δc are computed from un−1
τ and T n−1τ thanks to423
Eq. (24) in order to define the velocity and temperature profiles given by Eqs. (17)424
and (18).425
(d) These profiles are used to compute the right-hand side terms of Eqs. (25a)426
and (25b). Integral terms Ln−1f are approximated with a Gauss-Legendre quadra-427
ture. Note that the density profile is determined from the temperature profile428
using the perfect gas relation Eq. (9e). Convective terms Mn−1x and Mn−1
Txare429
computed using a second-order centered scheme. Knowing the form of the ve-430
locity and temperature profiles, their wall-normal derivatives can be analytically431
derived. They are written in Appendix B. Then, turbulent viscosity µn−1t and432
thermal conductivity λn−1t are given by Eqs. (13) and (14).433
21
(e) The wall fluxes at current time step unτ and T nτ are determined in such a way434
that the left-hand sides of Eqs. (25a) and (25b) equal their right-hand sides using435
a Newton-Raphson algorithm. un−1τ and T n−1
τ are used as initial guess. For436
each evaluation, the scalars parameters An, AnT , δc and δν are computed and the437
corresponding velocity and temperature profiles are used to approximate integral438
terms Lnf thanks to a Gauss-Legendre quadrature.439
(f) Once the Newton-Raphson algorithm has converged, the wall friction modulus440 ∥∥∥τnw
∥∥∥ and wall heat flux φwn
are fed back to the LES solver. The wall friction441
vector components in the flow solver frame are finally determined by switching442
from the boundary layer to the flow solver frame as explained in subsection II B.443
Newton-Raphson
algorithm
un1 , Tn1
An1, An1T
un , T
n
i An, AnT
Ln , L
nuLn1
, Ln1u
InitialGuess
un1 , Tn1
An1, An1T
i
un1 , Tn1
An1, An1T
i
y1
Mn1x
, Mn1T
x
eun11k
, eun1k
...
Ln1u , Ln1
u , Ln1u2
eun11k
, epn11 ... eun1
1k, epn1
1 ...
Ln1u , Ln1
u , Ln1u2
FIG. 3: Scheme of the iWMLES resolution algorithm.
IV. APPLICATION TO BI-PERIODIC PLANE CHANNEL FLOWS444
The integral wall-model for compressible flows described in Section III is assessed on445
plane channel flows and results are compared to available DNS data. After presenting the446
selected DNS configurations, the numerical setup used to launch WMLES computation is447
described. Finally the results on quasi-incompressible and supersonic cases are presented.448
A. Test cases449
In order to evaluate the capacity of the integral wall-model to accurately predict wall450
fluxes, several DNS of bi-periodic plane channel flow are considered as summarized in the451
22
Table I.452
For incompressible DNS, several friction Reynolds number Reτ are available (102070,453
200371, 417972 and 518673). For the Reτ = 1020 DNS, a temperature gradient is imposed by454
a constant and uniform heat-flux at walls. To reproduce this DNS in WMLES a constant455
wall temperature Tw is applied and the flow is forced such that Tb/Tw = 1.1 with Tb =456
∫ 2h
0ρuT dy /
∫ 2h
0ρu dy the bulk temperature and h the half-height channel. In the other457
DNS, the wall is considered adiabatic. For the WMLES simulations, a small temperature458
gradient Tb/Tw = 1.1 will also be applied in order to assess the accuracy of the wall heat flux459
computation. The adiabatic DNS friction coefficients are still considered as reference. The460
reference Nusselt number Nu = 2hφw/(λ(Tw − Tb)) is computed from Kays’s correlation74:461
Nu =0.023Re0.8
f Pr
0.88 + 2.03(Pr 2/3 − 0.78)Re−0.1f
(27)
with Ref = 2hρfub/µf the Reynolds number based on the film density ρf = 2ρwρb/(ρw+ρb),462
the film viscosity µf computed from the Sutherland’s law and the film temperature Tf =463
(Tb + Tw)/2. ρb =∫ 2h
0ρ dy/(2h) is the bulk density.464
For the compressible DNS under consideration, Reτ is equal to 1015, 663 or 972. At Reτ =465
1015, the Mach number Mb = ub/√γrTw based on the bulk velocity ub =
∫ 2h
0ρu dy/(2hρb)466
is equal to 1.5. In the other two cases, Mb = 1.7. For all cases, only non-adiabatic boundary467
conditions exist.468
For all plane channel flows simulations, the flow is forced to prevent the decrease of469
kinetic energy due to dissipation. The compressible DNS are reproduced here by applying470
a constant wall temperature Tw and forcing the bulk temperature Tb to get the same ratio471
Tb/Tw as in the corresponding DNS.472
In all cases, it has to be noted that the friction Reynolds number is an output of the473
computation as it depends on the friction velocity uτ . It is the bulk Reynolds number Reb =474
ρbubh/µb which defines the flow where µb is the bulk viscosity obtained from Sutherland’s law475
and the bulk temperature Tb. The wall heat flux coefficient Bq is defined by Bq = φwρwcpuτTw
476
and the friction Mach number Mτ by Mτ = uτ/√γrTw.477
23
TABLE I: Plane channel flow DNS considered
DNS (Nx, Ny, Nz) Boundary
conditions
Mb Reτ Reb Mτ Tb/Tw −Bq / Nu
Abe, Kawamura, and
Matsuo 70
(2048, 448, 1536) uniform heat-flux - 1020 20721 - . . . . . . / 83.0
Hoyas and Jimenez 71 (6144, 633, 4608) adiabatic - 2003 43590 - - -
Lozano-Duran and
Jimenez 72
(-, 1081, -) adiabatic - 4179 98302 - - -
Lee and Moser 73 (10240, 1536, 7680) adiabatic - 5186 124862 - - -
Modesti and
Pirozzoli 75(2048, 512, 1024) isothermal 1.5 1015 17000 0.065 1.35 0.038 / 123.5
Trettel and Larsson 76 (800, 246, 400) isothermal 1.7 663 10000 0.077 1.45 0.053 / 68.9
Trettel and Larsson 76 (896, 384, 480) isothermal 1.7 972 15500 0.073 1.45 0.050 / 112.7
B. Simulation set ups478
The cell-centered finite volume flow solver elsA-ONERA described in subsection II A is479
used to carry out all the simulations. In all adiabatic WMLES cases, the Prandtl number Pr480
is set to 0.72. In the other cases Pr is chosen accordingly to the DNS values: at Reτ = 1020481
Pr is set at 0.71 and in all supersonic cases Pr = 0.7. For all cases, the turbulent Prandtl482
number Pr t is assumed constant and equal to 0.9. In fact, using Germano-like dynamic483
calibration, one can observe that this parameter is not constant and is spatially varying. In484
practice, it appears to have a minor influence only77,78.485
The computational domain’s size is defined by (x, y, z) ∈ Ω = [0 ; 2πδ]× [0 ; 2δ]× [0 ;πδ]486
with δ the boundary layer thickness which is equal to the channel half-height h and x, y, and487
z respectively the coordinates in the longitudinal, wall-normal and spanwise directions. For488
high-Reynolds number flows like those considered in this study, the computational domain489
size is big enough to capture turbulent eddies and has been used in previous studies23,56,79.490
Two uniform grids G1 and G2 are considered: 25 × 21 × 21 and 49 × 41 × 41. On the491
former mesh, the first off-wall cell is located at y1 = 0.05δ while on the latter y1 = 0.025δ.492
24
Since cells size are approximatively given by ∆x = 5.2y1 and ∆z = πy1, cell aspect-ratio493
follows the outer-layer turbulent structure’s size. All cases parameters are summarized in494
tables II, III and IV.495
In order to enforce the correct mass flow rate and bulk temperature, energy is injected
in the domain through two volumic sources terms Sx and Qx which are respectively added
to the streamwise momentum and energy equations:
Sx(t) = (ρu)btarget −1
‖Ω‖
∫
Ω
ρu dΩ (28)
Qx(t) = Tbtarget −∫
ΩρuT dΩ∫
Ωρu dΩ
+ uSx (29)
Sx aims to force the flow to reach a given bulk Reynolds number by imposing the correct mass496
flow rate (ρu)btarget while Qx forces the bulk temperature. These two terms are computed497
at each iteration of the flow solver. Constant parts can be added to Sx and Qx in order498
to accelerate flow convergence80. The source terms are also taken into account in iWMLES499
equations (25a) and (25b). Moreover numerical experiments showed that identical results500
are obtained when Mx and MTx are taken into account if the wall-model input data are501
filtered in time49,81. Indeed in plane channel flow convective terms are statistically zero but502
due to the unsteady nature of LES, non-filtered Mx and MTx terms are always oscillating503
even after reaching flow convergence. In order to reduce the computational cost of the504
simulations, convective terms are neglected in the following.505
The initial condition is built from a steady state based on a power-law velocity profile:506
∀(x, y, z) ∈ Ω,
u(x, y, z, t = 0) =8
7ubtarget
(1− |1− y
h|)
v(x, y, z, t = 0) = 0
w(x, y, z, t = 0) = 0
T (x, y, z, t = 0) = Tw
(30)
which is perturbed with an additional random noise applied to the velocity field. The507
noise maximum amplitude is equal to 10% in the longitudinal direction and 5% in the508
other directions of the bulk velocity. It can be noted that random noise is enough to force509
turbulent transition of the flow because coarse grids are considered in this study. For finer510
grids as encountered in resolved LES, vortex rings are added to the initial flow to mimic511
turbulent structures56,82. However, WMLES grids are not fine enough to discretize these512
kind of structures.513
25
TABLE II: Adiabatic quasi-incompressible plane channel flow cases in WMLES
Case name (Nx, Ny, Nz) Mb Reτ Reb ∆+x y+
1 (at cell) ∆+z
G1 M0.2 R2003 adia (25, 21, 21) 0.2 2003 43590 524 100 315
G2 M0.2 R2003 adia (49, 41, 41) 0.2 2003 43590 262 50 157
G1 M0.2 R4179 adia (25, 21, 21) 0.2 4179 98302 1094 209 656
G2 M0.2 R4179 adia (49, 41, 41) 0.2 4179 98302 547 104 328
G1 M0.2 R5186 adia (25, 21, 21) 0.2 5186 124862 1358 259 815
G2 M0.2 R5186 adia (49, 41, 41) 0.2 5186 124862 679 130 407
TABLE III: Isothermal quasi-incompressible plane channel flow cases in WMLES
Case name (Nx, Ny, Nz) Mb Reτ Reb Tw/Tb ∆+x y+
1 (at cell) ∆+z
G1 M0.2 R1020 (25, 21, 21) 0.2 1020 20721 1.1 267 51 160
G2 M0.2 R1020 (49, 41, 41) 0.2 1020 20721 1.1 134 26 80
G1 M0.2 R2003 (25, 21, 21) 0.2 2003 43590 1.1 524 100 315
G2 M0.2 R2003 (49, 41, 41) 0.2 2003 43590 1.1 262 50 157
G1 M0.2 R4179 (25, 21, 21) 0.2 4179 98302 1.1 1094 209 656
G2 M0.2 R4179 (49, 41, 41) 0.2 4179 98302 1.1 547 104 328
G1 M0.2 R5186 (25, 21, 21) 0.2 5186 124862 1.1 1358 259 815
G2 M0.2 R5186 (49, 41, 41) 0.2 5186 124862 1.1 679 130 407
TABLE IV: Isothermal supersonic plane channel flow cases in WMLES
Case name (Nx, Ny, Nz) Mb Reτ Reb Tb/Tw ∆+x y+
1 (at cell) ∆+z
G1 M1.5 R1015 (25, 21, 21) 1.5 1015 17000 1.35 266 51 159
G2 M1.5 R1015 (49, 41, 41) 1.5 1015 17000 1.35 133 25 80
G1 M1.7 R663 (25, 21, 21) 1.7 663 10000 1.45 174 33 104
G2 M1.7 R663 (49, 41, 41) 1.7 663 10000 1.45 87 17 52
G1 M1.7 R972 (25, 21, 21) 1.7 972 15500 1.45 254 49 153
G2 M1.7 R972 (49, 41, 41) 1.7 972 15500 1.45 127 24 76
26
For all simulations, flow convergence toward a statistically steady state is assumed to be514
reached once the kinetic energy’s variation in the whole domain does not exceed 0.01% of515
its mean value. Then, statistics are obtained by time averaging the flow over 200× 2πδ/ub516
periods through the channel. Space averaging along longitudinal and spanwise directions517
are also done before comparing the mean profiles with DNS data.518
C. Results for quasi-incompressible flows519
Before evaluating the capacity of iWMLES to take into account complex physical phe-520
nomena such as compressibility effects, it is required to check whether iWMLES is able to521
recover standard results where aWMLES is known to be accurate. As explained in subsec-522
tion IV A, four incompressible DNS are considered. Besides a small temperature gradient523
is also applied on all adiabatic cases in order to evaluate iWMLES on quasi-incompressible524
isothermal flows. Therefore, seven quasi-incompressible cases in terms of Reynolds number525
and wall boundary conditions are computed with aWMLES and iWMLES. All numerical526
and physical cases parameters are summarized in tables II and III.527
Mean profiles in wall units for adiabatic and isothermal cases are respectively shown in528
Figs. 4 and 6. In all cases, iWMLES recovers the logarithmic law of the wall and is in agree-529
ment with the DNS profiles, which means the corrective terms added in the parametrized530
velocity and temperature profiles cancel out for different Reynolds numbers and grids. Sim-531
ilar profiles are obtained with aWMLES. No log-layer mismatch is observed unlike other532
hybrid RANS/LES methods10,13, which justifies the numerical setup employed in this study.533
Velocity and temperature fluctuations at Reτ = 1020 and 5186 are shown in Figs. 5, 7534
and 8. In all cases, aWMLES and iWMLES predict almost identical turbulent fluctuations535
whose overall level are either in good agreement with DNS when it is available or coherent536
otherwise. The near-wall peak on streamwise velocity is not captured due to the coarse537
grid used and seems to start to be captured when the grid is refined. The only major538
difference between iWMLES and aWMLES occurs in the temperature fluctuations profiles539
at the first off-wall cell. As only DNS data at Reτ = 1020 is available, it cannot be said540
whether iWMLES’s temperature fluctuation estimation at the first cell is more accurate541
than aWMLES.542
Relative errors on the bulk friction coefficient and the Nusselt number are shown in543
27
Table V. As observed with the mean profiles, accurate wall friction fluxes are predicted by544
both iWMLES and aWMLES. In adiabatic cases a maximum error of 3.1% is obtained. In545
isothermal cases, errors on the friction coefficient increase with the Reynolds number with546
maximum errors of 9.1% for aWMLES and 8.5% for iWMLES. For Reτ ≥ 2003, the reference547
wall friction is taken from incompressible adiabatic DNS. In WMLES, as a compressible548
code is considered and due to the imposed temperature gradient, a density gradient exists.549
Therefore, the assumption that the DNS friction coefficient is still valid may not hold when550
the Reynolds number increases. The largest errors on the Nusselt number occur at the551
lowest Reynolds number. This result may be due to the position of the first off-wall cell552
approaching the buffer layer. At higher Reynolds number cases, iWMLES and aWMLES553
are both in agreement with Kays’s correlation.554
To sum up, iWMLES is able to recover the standard logarithmic laws for both the velocity555
and temperature profiles. Velocity and temperature profiles (mean and fluctuations) are in556
good agreement with adiabatic and isothermal DNS data. It is interesting to note that557
WMLES using either aWMLES or iWMLES predict nearly identical fluctuations even if558
they are based on two different approaches. From a numerical point-of-view, differences559
could be observed using high-order schemes. From a physical point-of-view, velocity and560
temperature fluctuations are expected to be much more dependent on the features of the561
numerical methods (convective flux scheme, subgrid-scale model, etc.) used than the wall-562
model. Indeed, this result is supported by recent studies of Cossu and Hwang 83 . They563
show that outer layer large-scale structures of wall-bounded turbulent flows can self-sustain564
without near-wall smaller-scale, by extracting energy from the mean flow. In WMLES, near-565
walls grids are too coarse to capture near-walls streaks but only large turbulent structures.566
Therefore, turbulence fluctuations would depend mostly on the resolution of these large-scale567
structures. Unlike near-wall small-scale structures, they are not expected to depend heavily568
on wall fluxes. This could explain the relative independance of the velocity fluctuations with569
respect to the wall-model used.570
D. Results for isothermal supersonic flows571
More discriminant test cases, namely supersonic channel flows, are now considered. Three572
different cases in terms of Reynolds and Mach numbers are computed as shown in Table IV.573
28
TABLE V: Relative errors in % on the bulk friction coefficient Cfb = τw12ρbu
2b
and the Nusselt
number Nu = 2hφwλ(Tw−Tb)
for quasi-incompressible plane channel flow cases in WMLES. *:
reference Nusselt number is computed from Kays’s correlation
Case Wall-model ∆(Cfb) in % ∆(Nu) in %
G1 M0.2 R2003 adia aWMLES 0.3 -
G2 M0.2 R2003 adia aWMLES 0.7 -
G1 M0.2 R4179 adia aWMLES 1.3 -
G2 M0.2 R4179 adia aWMLES 2.5 -
G1 M0.2 R5186 adia aWMLES 2.3 -
G2 M0.2 R5186 adia aWMLES 3.1 -
G1 M0.2 R2003 adia iWMLES −0.3 -
G2 M0.2 R2003 adia iWMLES −0.1 -
G1 M0.2 R4179 adia iWMLES 2.0 -
G2 M0.2 R4179 adia iWMLES 1.4 -
G1 M0.2 R5186 adia iWMLES 2.5 -
G2 M0.2 R5186 adia iWMLES 2.3 -
G1 M0.2 R1020 aWMLES 3.8 13.7
G2 M0.2 R1020 aWMLES 5.1 14.6
G1 M0.2 R2003 aWMLES 5.3 −1.9∗
G2 M0.2 R2003 aWMLES 5.3 −1.3∗
G1 M0.2 R4179 aWMLES 7.1 −0.04∗
G2 M0.2 R4179 aWMLES 8.0 1.0∗
G1 M0.2 R5186 aWMLES 8.2 0.7∗
G2 M0.2 R5186 aWMLES 9.1 1.8∗
G1 M0.2 R1020 iWMLES 1.2 7.4
G2 M0.2 R1020 iWMLES 7.1 14.0
G1 M0.2 R2003 iWMLES 4.7 −4.5∗
G2 M0.2 R2003 iWMLES 6.5 −2.2∗
G1 M0.2 R4179 iWMLES 7.7 −1.8∗
G2 M0.2 R4179 iWMLES 6.8 −2.2∗
G1 M0.2 R5186 iWMLES 7.9 −1.5∗
G2 M0.2 R5186 iWMLES 8.5 −0.7∗
29
0
10
20
30
40
50
60
70
80
1 10 100 1000 10000
u+
y+
aWMLESiWMLES
DNSReichardt law
FIG. 4: Mean velocity profiles in wall units for adiabatic quasi-incompressible cases.
Profiles are shifted by multiple of 10. From bottom to top: G1 M0.2 R2003 adia,
G2 M0.2 R2003 adia, G1 M0.2 R4179 adia, G2 M0.2 R4179 adia, G1 M0.2 R5186 adia,
G2 M0.2 R5186 adia.
In all cases a constant wall temperature is imposed in order to cool the flow and evacuate574
the injected energy through forcing terms as pointed out by Trettel and Larsson 76 . Usual575
logarithmic laws of the wall are not expected to hold on such flows. Thanks to its correc-576
tion terms and computation of a weak solution of the vertically integrated boundary layer577
equations, iWMLES is expected to be more accurate than aWMLES.578
Mean velocity profiles in wall units are shown in Fig. 9. As expected, aWMLES imposes579
the first off-wall cell to follow Reichardt’s law which is incorrect here. Due to compressibility580
30
0
1
2
3
0 1
u′ i+
y/h
aWMLESiWMLES
DNS
(a) Case G1 M0.2 R5186 adia
0
1
2
3
0 1
u′ i+
y/h
aWMLESiWMLES
DNS
(b) Case G2 M0.2 R5186 adia
FIG. 5: Velocity fluctuations for adiabatic cases. For each case, from bottom to top: v′+,
w′+ and u′+.
0
10
20
30
40
50
60
70
80
90
100
1 10 100 1000 10000
u+
y+
aWMLESiWMLES
DNSReichardt law
0
10
20
30
40
50
60
70
80
90
100
1 10 100 1000 10000
T+
y+
aWMLESiWMLESDNSKader law
FIG. 6: Mean velocity and temperature profiles in wall units for isothermal
quasi-incompressible cases. Profiles are shifted by multiple of 10. From bottom to top:
G1 M0.2 R1020, G2 M0.2 R1020, G1 M0.2 R2003, G2 M0.2 R2003, G1 M0.2 R4179,
G2 M0.2 R4179, G1 M0.2 R5186, G2 M0.2 R5186.
31
0
1
2
3
0 1
u′ i+
y/h
aWMLESiWMLES
DNS
(a) Case G1 M0.2 R1020 isot
0
1
2
3
0 1
u′ i+
y/h
aWMLESiWMLES
DNS
(b) Case G2 M0.2 R1020 isot
0
1
2
3
0 1
u′ i+
y/h
aWMLESiWMLES
(c) Case G1 M0.2 R5186 isot
0
1
2
3
0 1
u′ i+
y/h
aWMLESiWMLES
(d) Case G2 M0.2 R5186 isot
FIG. 7: Velocity fluctuations for isothermal cases. For each case, from bottom to top: v′+,
w′+ and u′+.
effects, velocity profiles are shifted upwards. This effect has also been observed on boundary581
layer flows84. Therefore, the friction velocity is overestimated and the whole velocity profiles582
in wall units are below the DNS profiles. On the contrary, iWMLES predicts correctly the583
wall friction. In all cases the first off-wall point is in agreement with the DNS data as well584
as the whole profile. Only the second and third off-wall cells are shifted under the DNS585
velocity profiles for the case G2 M1.7 663 (Reτ = 663 on the finer grid G2). This can be due586
to the position of the wall-model interface which is in the buffer layer.587
Likewise, mean temperature profiles in wall units for the aWMLES are not in agreement588
with the DNS profiles as shown in Fig. 9. Wall heat flux is systematically underestimated as589
the first off-wall cell is above the DNS. Temperature profiles are known to be more affected590
than velocity profiles64,85. The iWMLES model predicts more accurate wall heat flux than591
aWMLES and more coherent temperature profiles are obtained. As observed for the velocity592
in the case G2 M1.7 663, the temperature profile does not follow the usual trend observed593
in other cases: the wall heat flux is slightly overestimated in iWMLES and the temperature594
32
0
1
2
3
0 1
T′+
y/h
aWMLESiWMLES
DNS
(a) Case G1 M0.2 R1020 isot
0
1
2
3
0 1
T′+
y/h
aWMLESiWMLES
DNS
(b) Case G2 M0.2 R1020 isot
0
1
2
3
0 1
T′+
y/h
aWMLESiWMLES
(c) Case G1 M0.2 R5186 isot
0
1
2
3
0 1
T′+
y/h
aWMLESiWMLES
(d) Case G2 M0.2 R5186 isot
FIG. 8: Temperature fluctuations for isothermal cases.
profile is below the DNS reference.595
Observations on mean profiles are confirmed by the measured errors on the friction Mach596
number Mτ and the heat flux coefficient Bq in Table VI. Indeed larger errors are obtained597
with aWMLES compare to iWMLES. The maximum error on the friction Mach number is598
of 6.7% with the former and only 2.7% with the latter. For the heat flux coefficient, larger599
errors are computed as it is underestimated by at most 21.3% with aWMLES and 15.2%600
with iWMLES. Here, the aWMLES model shows its limits as the standard logarithmic laws601
are no longer valid on these kind of flows. In all cases, iWMLES predicts more accurately602
wall fluxes than aWMLES.603
Velocity and temperature fluctuations for Reτ = 972 are shown respectively in Figs. 10604
and 11. Other cases are not displayed for conciseness, as equivalent results and conclusions605
are obtained. Despite predicting different mean profiles, velocity fluctuations are very similar606
between iWMLES and aWMLES even if the former takes into account more physics than607
the latter in order to compute wall fluxes. Indeed, wall-models implemented here influence608
only the mean profiles and does not interact directly with the second order fluctuations609
33
terms. Like in quasi-incompressible flow cases, overall values of velocity fluctuations are in610
good agreement with the DNS data, except the first rows of cells above the wall where the611
fluctuations are overestimated. Compared to quasi-incompressible flows DNS, the near-wall612
peak is more spread out in supersonic flows. This can explain the overestimation of the613
fluctuations near walls. Indeed when the grid is refined, near-walls peak is better captured614
and better agreement with the DNS data is obtained.615
Finally, in contrast to the velocity fluctuations, the temperature fluctuations are not616
similar between aWMLES and iWMLES. Indeed, temperature fluctuations with iWMLES617
are in correct agreement with the DNS profiles on the coarser grid while aWMLES strongly618
overestimates it. Moreover, when the grid is refined, better agreement is obtained as the619
first off-wall cell is located at the near-wall peak position and iWMLES approach predicts620
temperature fluctuations close to DNS reference.621
To sum up, iWMLES has a wider domain of validity and is able to compute more accurate622
wall fluxes than aWMLES, especially the wall friction. Better mean profiles are obtained623
with iWMLES. Furthermore, as shown in subsection III E, iWMLES cost is only slightly624
more expensive compare to aWMLES.625
34
0
10
20
30
40
50
60
70
80
1 10 100 1000
u+
y+
aWMLESiWMLES
DNSReichardt law
0
10
20
30
40
50
60
70
1 10 100 1000
T+
y+
aWMLESiWMLESDNSKader law
FIG. 9: Mean velocity and temperature profiles in wall units for isothermal supersonic
cases. Profiles are shifted by multiple of 10. From bottom to top: G1 M1.5 R1015,
G2 M1.5 R1015, G1 M1.7 R663, G2 M1.7 R663, G1 M1.7 R972, G2 M1.7 R972.
0
1
2
3
4
0 1
u′ i+
y/h
aWMLESiWMLES
DNS
(a) Case G1 M1.7 R972 isot
0
1
2
3
4
0 1
u′ i+
y/h
aWMLESiWMLES
DNS
(b) Case G2 M1.7 R972 isot
FIG. 10: Velocity fluctuations for isothermal supersonic cases. For each case, from bottom
to top: v′+, w′+ and u′+.
35
0
0.5
1
1.5
2
0 1
T′+
y/h
aWMLESiWMLES
DNS
(a) Case G1 M1.7 R972 isot
0
0.5
1
1.5
2
0 1
T′+
y/h
aWMLESiWMLES
DNS
(b) Case G2 M1.7 R972 isot
FIG. 11: Temperature fluctuations for isothermal supersonic cases.
TABLE VI: Relative errors in % on the friction Mach number Mτ = uτcw
(cw is the speed of
sound) and the wall heat flux coefficient Bq = φwρwcpuτTw
= − TτTw
for supersonic plane channel
flow cases in WMLES
Case Wall-model ∆(Mτ ) in % ∆(Bq) in %
G1 M1.5 R1015 aWMLES 6.7 −20.5
G2 M1.5 R1015 aWMLES 5.9 −14.3
G1 M1.7 R663 aWMLES 3.9 −21.3
G2 M1.7 R663 aWMLES 4.1 −12.7
G1 M1.7 R972 aWMLES 6.1 −21.1
G2 M1.7 R972 aWMLES 4.9 −15.2
G1 M1.5 R1015 iWMLES 0.2 −13.4
G2 M1.5 R1015 iWMLES 1.8 −6.8
G1 M1.7 R663 iWMLES −2.7 −14.2
G2 M1.7 R663 iWMLES 0.5 −1.8
G1 M1.7 R972 iWMLES −2.2 −15.2
G2 M1.7 R972 iWMLES −0.2 −8.2
36
V. CONCLUDING REMARKS626
In this study, an attempt is made to develop a valid wall-model for a typical turboma-627
chinery flow range, i.e flows with Mach numbers up to about 1.5, Reynolds numbers between628
106 and 107 and temperature gradients in the order of ten to hundred Kelvin. For this pur-629
pose, a compressible extension of the integral wall-model approach originally introduced for630
incompressible flows has been proposed, namely the iWMLES. To this end, both the mean631
velocity and the mean temperature profile have been parametrized using logarithmic laws of632
the wall with additional correction terms. These profiles are coupled to integral relations for633
compressible thin boundary layers in order to determine unknowns parameters. Therefore,634
iWMLES takes into account more physics than standard equilibrium wall-models. Besides,635
the computational cost is reduced by a factor about 100 compared to wall-models based on636
the numerical resolution of thin boundary layer equations since only a local simple scalar637
system is solved. An analytical wall-model, namely the aWMLES, based on the standard638
logarithmic laws of the wall, is also implemented and used as a basis of comparison.639
iWMLES is validated on both isothermal and adiabatic wall cases, and is observed to640
accurately predict both subsonic and supersonic flows. Indeed, in subsonic flows, mean641
velocity and temperature profiles are in agreement with the DNS reference data, as well642
as the turbulent fluctuations profiles. iWMLES is therefore able to recover the standard643
logarithmic laws of the wall. In supersonic flows, as expected, aWMLES fails to predict644
the mean flow. On the contrary, iWMLES, by estimating more accurately the wall fluxes,645
is able to obtain results in agreement with the DNS data. Interestingly, both wall-models646
considered in this study predict nearly identical velocity fluctuations. This result indicates647
that the turbulent fluctuations are weakly dependant on the wall-model.648
To go further, and before applying iWMLES to turbomachinery applications, it could be649
interesting to apply iWMLES in flows with: (i) strong non-equilibrium effects ; (ii) strong650
compressibility effects ; (iii) complex wall boundaries. These effects could be studied by651
simulating an oblique shock/boundary-layer interaction, hypersonic flows or wavy walls86,87.652
For the latter case, taking into account curvature effects could be considered. However, this653
case has, for the moment, not been simulated at a Reynolds number high enough for WM-654
LES. Furthermore, DNS of spatially evolving supersonic boundary layers with isothermal655
walls88–92 have been recently performed. Friction Reynolds numbers reached are increasing656
37
and these cases could later be used as reference cases in order to evaluate wall-models for657
heat transfer prediction. Nonetheless, iWMLES, by taking into account more physics, is658
able to extend the domain of validity observed with a standard analytical wall-model with659
similar computational cost.660
ACKNOWLEDGMENTS661
The authors acknowledge A. Trettel and J. Larsson for kindly providing DNS data of662
their supersonic channel flows. This work was supported by the Association Nationale de la663
Recherche et de la Technologie (ANRT ; Grant No. 2015/0193).664
Appendix A: Jacobian matrix of the iWMLES equations665
As explained in subsection III E, a Newton-Raphson algorithm is used to solve the system666
of equations (25). In this appendix the corresponding Jacobian matrix is detailed in the case667
of an isothermal boundary condition. If the wall is considered to be adiabatic, the Jacobian668
matrix is reduced to a one-dimensional function.669
Let f and g be two functions defined by:
f(unτ ;T nτ ) = Lnρu − un−11‖
Lnρ − Ln−1ρu (1− un−1
1‖)−∆t(−Mn−1
x + τ1n−1 − τwn−1) (A1)
g(unτ ;T nτ ) = cpTn−11 Lnρ −
(cpr− 1)y1(pn1 − pn−1
1 )− cpT n−11 Ln−1
ρ (A2)
+ ∆t(−Mn−1Tx
+ Ln−1
τ ∂u∂y
− φ1n−1 + φw
n−1)
Variables at time step n − 1 are known and can be considered as constant scalars of the670
equation system as well as the pressure pn which is an input from the LES to the wall-model671
equations. Friction velocity unτ and temperature T nτ at time step n are the unknown. The672
system of equations (25) can be simply written as:673
f(unτ ;T nτ ) = 0
g(unτ ;T nτ ) = 0(A3)
and the corresponding Jacobian matrix J is by definition:674
J =
∂f∂unτ
∂f∂Tnτ
∂g∂unτ
∂g∂Tnτ
(A4)
38
J components are given by:
∂f
∂unτ=∂Lnρu∂unτ
− un−11‖
∂Lnρ∂unτ
(A5a)
∂f
∂T nτ=∂Lnρu∂T nτ
− un−11‖
∂Lnρ∂T nτ
(A5b)
∂g
∂unτ= cpT
n−11
∂Lnρ∂unτ
(A5c)
∂g
∂T nτ= cpT
n−11
∂Lnρ∂T nτ
(A5d)
675
Thus, J is completely defined by the derivatives of integral terms Lnρ and Lnρu with respect
to unτ and T nτ . As the fluid is supposed to be a perfect gas, density and temperature profile
are linked with pressure ρn(y) = pn1/(rTn(y)) and:
∂Lnρ∂unτ
=pn1r
∫ y1
0
−(
1
T n
)2∂T n
∂unτdy (A6a)
∂Lnρ∂T nτ
=pn1r
∫ y1
0
−(
1
T n
)2∂T n
∂T nτdy (A6b)
∂Lnρu∂unτ
=pn1r
∫ y1
0
∂un
∂unτT n − un ∂Tn
∂unτ
T n2 dy (A6c)
∂Lnρu∂T nτ
=pn1r
∫ y1
0
∂un
∂TnτT n − un ∂Tn
∂Tnτ
T n2 dy (A6d)
676
Temperature T n and velocity profile un between the wall assumed to be located at y = 0
and the first off-wall point y1 are given by Eqs. (17) and (18), so:
∂un
∂unτ=un
unτ+ unτ
(y+n
unτ (1 + κy+n)+ (B − 1
κlnκ)
y+n
11unτ
[exp(−y
+n
11)
+(y+n
3− 1) exp(−y
+n
3)
]+ An
√y+n
y+n
δνunτexp(−y
+n
δν)
+∂An
∂unτ
√y+n(1− exp(−y
+n
δν)) +
An√y+n
2unτ(1− exp(−y
+n
δν))
)(A7a)
∂un
∂T nτ= 0 (A7b)
39
and:
1
T nτ
∂T n
∂unτ= Pry+n exp(−Γ)(
1
unτ− ∂Γ
∂unτ) +
∂Γ
∂unτ
exp(− 1Γ)
Γ2(
1
κtln(1 + y+n) + β)
+ exp(− 1
Γ)
y+n
κtunτ (1 + y+n)+ AT
y+n2
δcunτexp(−y
+n
δc)
+ (∂AT∂unτ
+ATunτ
)y+n(1− exp(−y+n
δc))
(A8a)
∂T n
∂T nτ=T n − TwT nτ
+ Tτ∂AT∂Tτ
y+n(1− exp(−y+n
δc)) (A8b)
with κ, κt, β and Γ defined in Eq. (8). So:677
∂Γ
∂unτ= 10−2(Pry+n)4 4 + 15Pr 3y+n
unτ (1 + 5Pr 3y+n)2(A9)
The derivatives of An and AnT are obtained from Eq. (24):
∂An
∂unτ=
[−un1‖unτ
2− y+
1n
unτ (1 + κy+1 )− (B − 1
κlnκ)
y+1n
11unτ
(exp(−y
+1n
11)
+ exp(−y+1n
3)(y+
1n
3− 1)
)]×√y+
1n(1− exp(−y
+1n
δν))
−[(2y+
1n
δν− 1) exp(−y
+1n
δν) + 1
]× Any
+1n
2unτ(1− exp(−y
+1n
δν))
/(y+
1n(1− exp(−y
+1n
δν))2)
(A10a)
∂AnT∂unτ
=
[−Pry+
1n
exp(−Γ1)(1
unτ− ∂Γ1
∂unτ)− exp(− 1
Γ1
)
(y+
1n
κtunτ (1 + y+1n)
+ (1
κtln(1 + y+
1n) + β)
1
Γ21
∂Γ1
∂unτ
)]× y+
1n(1− exp(−y
+1n
δc))
−[(y+
1n
δc− 1) exp(−y
+1n
δc) + 1
]× AnT
y+1n2
unτ(1− exp(−y
+1n
δc))
/(y+
1n(1− exp(−y
+1n
δc)))2
(A10b)
∂AnT∂T nτ
= − T1
n − TwT nτ
2y+1n(1− exp(−y+1
n
δc))
(A10c)
where Γ1 = Γ(y1). Note that ∂An
∂Tnτ= 0.678
Therefore, using Eqs. (A7), (A8) and (A10) the integrated terms in Eq. (A6) and so679
the Jacobian matrix components (A5) can be expressed analytically. From an initial guess680
(unτ0 , Tnτ0
), the Newton-Raphson algorithm consists in computing a solution to Eq. (A3) by681
40
iteration:682
unτi+1= unτi +
∂f∂Tnτ
g − f ∂g∂Tnτ
∂f∂unτ
∂g∂Tnτ− ∂f
∂Tnτ
∂g∂unτ
T nτi+1= T nτi −
f + ∂f∂unτ
( ∂f∂Tnτ
g − f ∂g∂Tnτ
)
∂f∂Tnτ
( ∂f∂unτ
∂g∂Tnτ− ∂f
∂Tnτ
∂g∂unτ
)
(A11)
As explained in subsection III E, the wall friction velocity and the wall friction temper-683
ature at previous time step are taken as initial guess. At each iteration of the algorithm,684
a Gauss-Legendre quadrature is used to compute the Jacobian matrix components. The685
algorithm is stopped once f(unτ ;T nτ ) and g(unτ ;T nτ ) are both below a given threshold. In this686
study, the thresold is set at 10−13.687
In the case of an adiabatic boundary condition, Tτ is zero. Therefore, ∂f∂unτ
is the only688
component of J and:689
∂T n
∂unτ= −T n1 Pr 1/3(γ − 1)Mn2
1‖
un
un2
1‖
∂un
∂unτ(A12)
where ∂un
∂unτis given by Eq. (A7a).690
Appendix B: Wall-normal derivatives of the parametrized profiles691
In iWMLES equations (25), flux values at the first off-wall point τ 1 = (µ+µt|y1)∂u∂y |y1 and
φ1 = −(λ+λt|y1)∂T∂y |y1 , as well as the wall-normal integral of τ∂u/∂y, are required. Turbulent
viscosity µt and thermal conductivity λt are modeled by mixing-length models (see Eqs. (13)
and (14)). Therefore, wall-normal derivatives of the velocity and temperature are needed.
As explained in subsection III E, they are obtained by derivation of the parametrized profiles
(17) and (18). Their expressions are given here:
1
uτ
∂u
∂y=
y+
y(1 + κy+)+ (B − 1
κlnκ)
y+
11y
(exp(−y
+
11) + (
y+
3− 1) exp(−y
+
3)
)
+A√y+
2y
(1 + (2
y+
δν− 1) exp(−y
+
δν)
) (B1a)
1
Tτ
∂T
∂y= Pry+ exp(−Γ)(
1
y− ∂Γ
∂y) + exp(− 1
Γ)
(y+
κty(1 + y+)
+(1
κtln(1 + y+) + β)
1
Γ2
∂Γ
∂y
)+ AT
y+
y
(1 + (
y+
δc− 1) exp(−y
+
δc)
) (B1b)
with:692
∂Γ
∂y= 10−2(Pry+)4 4 + 15Pr 3y+
y(1 + 5Pr 3y+)2(B2)
41
REFERENCES693
1N. Gourdain, L. Y. M. Gicquel, and E. Collado, “RANS and LES for the heat transfer694
prediction in turbine guide vane,” Journal of Propulsion and Power 28, 423 (2012).695
2P. Sagaut, Large-eddy simulation for incompressible flows, 3rd edition, edited by Springer696
(Springer, 2005).697
3E. Garnier, N. Adams, and P. Sagaut, Large-eddy simulation for compressible flows , edited698
by Springer (Springer, 2009).699
4P. Sagaut and S. Deck, “Large eddy simulation for aerodynamics: status and perspec-700
tives,” Philosophical Transactions of the Royal Society A: Mathematical, Physical and701
Engineering Sciences 367, 2849 (2009).702
5A. J. Smits, B. J. McKeon, and I. Marusic, “High–reynolds number wall turbulence,”703
Annual Review of Fluid Mechanics 43, 353 (2011).704
6D. R. Chapman, “Computational aerodynamics development and outlook,” AIAA Journal705
17, 1292 (1979).706
7H. Choi and P. Moin, “Grid-point requirements for large eddy simulation: Chapman’s707
estimates revisited,” Physics of Fluids 24, 011702 (2012).708
8U. Piomelli and E. Balaras, “Wall-layer models for large-eddy simulations,” Annual Review709
of Fluid Mechanics 34, 349 (2002).710
9P. Sagaut, S. Deck, and M. Terracol, Multiscale and multiresolution approaches in turbu-711
lence, 2nd edition, edited by I. C. Press (Imperial College Press, 2013).712
10U. Piomelli, “Wall-layer models for large-eddy simulations,” Progress in Aerospace Sciences713
44, 437 (2008).714
11J. Frohlich and D. von Terzi, “Hybrid les/rans methods for the simulation of turbulent715
flows,” Progress in Aerospace Sciences 44, 349 (2008).716
12P. R. Spalart, “Detached-eddy simulation,” Annual Review of Fluid Mechanics 41, 181717
(2009).718
13J. Larsson, S. Kawai, J. Bodart, and I. Bermejo-Moreno, “Large eddy simulation with719
modeled wall-stress: recent progress and future directions,” Mechanical Engineering Re-720
views 3, 1 (2016).721
14S. T. Bose and G. I. Park, “Wall-modeled large-eddy simulation for complex turbulent722
flows,” Annual Review of Fluid Mechanics 50, 535 (2018).723
42
15S. Deck, N. Renard, R. Laraufie, and P. Sagaut, “Zonal detached eddy simulation (zdes)724
of a spatially developing flat plate turbulent boundary layer over the reynolds number725
range 3150 ≤ Reθ ≤ 14000,” Physics of Fluids 26, 025116 (2014).726
16S. Deck and N. Renard, “Improvements in zonal detached eddy simulation for wall modeled727
large-eddy simulation,” AIAA Journal 53, 3599 (2015).728
17P. Spalart and S. Allmaras, “A one-equation turbulence model for aerodynamic flows,”729
in 30th Aerospace Sciences Meeting and Exhibit (American Institute of Aeronautics and730
Astronautics (AIAA), 1992).731
18M. Strelets, “Detached eddy simulation of massively separated flows,” in 39th Aerospace732
Sciences Meeting and Exhibit (American Institute of Aeronautics and Astronautics, 2001).733
19F. Hamba, “Analysis of filtered navierstokes equation for hybrid RANS/LES simulation,”734
Physics of Fluids 23, 015108 (2011).735
20U. Piomelli, E. Balaras, H. Pasinato, K. D. Squires, and P. R. Spalart, “The inner-outer736
layer interface in large-eddy simulations with wall-layer models,” International Journal of737
Heat and Fluid Flow 24, 538 (2003).738
21A. Keating and U. Piomelli, “A dynamic stochastic forcing method as a wall-layer model739
for large-eddy simulation,” Journal of Turbulence 7, N12 (2006).740
22S. Kawai and J. Larsson, “Wall-modeling in large eddy simulation: Length scales, grid741
resolution, and accuracy,” Physics of Fluids 24, 015105 (2012).742
23S. Bocquet, P. Sagaut, and J. Jouhaud, “A compressible wall model for large-eddy sim-743
ulation with application to prediction of aerothermal quantities,” Physics of Fluids 24,744
065103 (2012).745
24J. Jimenez, “Cascades in wall-bounded turbulence,” Annual Review of Fluid Mechanics746
44, 27 (2012).747
25W. Cabot and P. Moin, “Approximate wall boundary conditions in the large-eddy simu-748
lation of high reynolds number flow,” Flow, Turbulence and Combustion 63, 269 (2000).749
26P. Wu and J. Meyers, “A constraint for the subgrid-scale stresses in the logarithmic region750
of high reynolds number turbulent boundary layers: A solution to the log-layer mismatch751
problem,” Physics of Fluids 25, 015104 (2013).752
27L. S. G. Kovasznay, V. Kibens, and R. F. Blackwelder, “Large-scale motion in the in-753
termittent region of a turbulent boundary layer,” Journal of Fluid Mechanics 41, 283754
(1970).755
43
28K. C. Kim and R. J. Adrian, “Very large-scale motion in the outer layer,” Physics of Fluids756
11, 417 (1999).757
29N. Hutchins and I. Marusic, “Evidence of very long meandering features in the logarithmic758
region of turbulent boundary layers,” Journal of Fluid Mechanics 579, 1 (2007).759
30G. I. Park and P. Moin, “Numerical aspects and implementation of a two-layer zonal760
wall model for les of compressible turbulent flows on unstructured meshes,” Journal of761
Computational Physics 305, 589 (2016).762
31G. Hoffmann and C. Benocci, “Approximate wall boundary conditions for large eddy763
simulations,” in Fluid Mechanics and Its Applications (Springer Netherlands, 1995) p.764
222.765
32E. Balaras and C. Benocci, “Two-layer approximate boundary conditions for large-eddy766
simulations,” AIAA Journal 34, 1111 (1996).767
33M. Wang and P. Moin, “Dynamic wall modeling for large-eddy simulation of complex768
turbulent flows,” Physics of Fluids 14, 2043 (2002).769
34S. Kawai and J. Larsson, “Dynamic non-equilibrium wall-modeling for large eddy simula-770
tion at high reynolds numbers,” Physics of Fluids 25, 015105 (2013).771
35G. I. Park and P. Moin, “An improved dynamic non-equilibrium wall-model for large eddy772
simulation,” Physics of Fluids 26, 015108 (2014).773
36S. Hickel, E. Touber, J. Bodart, and J. Larsson, “A parametrized non-equilibrium wall-774
model for large-eddy simulations,” (2012).775
37I. Bermejo-Moreno, L. Campo, J. Larsson, J. Bodart, D. Helmer, and J.-K. Eaton,776
“Confinement effects in shock wave/turbulent boundary layer interactions through wall-777
modelled large-eddy simulations,” Journal of Fluid Mechanics 758, 5 (2014).778
38X. I. A. Yang, J. Urzay, S. Bose, and P. Moin, “Aerodynamic heating in wall-modeled779
large-eddy simulation of high-speed flows,” AIAA Journal 56, 731 (2018).780
39A. Weber, H.-A. Schreiber, R. Fuchs, and W. Steinert, “3-d transonic flow in a compressor781
cascade with shock-induced corner stall,” Journal of Turbomachinery 124, 358 (2002).782
40C. Hah, “Large eddy simulation of transonic flow field in nasa rotor 37,” Tech. Rep. TM-783
2009-215627 (2009).784
41X. Ottavy, N. Courtiade, and N. Gourdain, “Experimental and computational methods785
for flow investigation in high-speed multistage compressor,” Journal of Propulsion and786
Power 28, 1141 (2012).787
44
42T. Leonard, L. Y. M. Gicquel, N. Gourdain, and F. Duchaine, “Steady/unsteady reynolds-788
averaged navier–stokes and large eddy simulations of a turbine blade at high subsonic outlet789
mach number,” Journal of Turbomachinery 137, 041001 (2014).790
43N. Gourdain, “Prediction of the unsteady turbulent flow in an axial compressor stage. part791
1: Comparison of unsteady RANS and LES with experiments,” Computers & Fluids 106,792
119 (2015).793
44J. Tyacke, P. Tucker, R. Jefferson-Loveday, N. R. Vadlamani, R. Watson, I. Naqavi, and794
X. Yang, “Large eddy simulation for turbines: Methodologies, cost and future outlooks,”795
Journal of Turbomachinery 136, 061009 (2014).796
45P. G. Tucker and J. C. Tyacke, “Eddy resolving simulations in aerospace in-797
vited paper (numerical fluid 2014),” Applied Mathematics and Computation (2015),798
10.1016/j.amc.2015.02.018.799
46T. Arts and M. Lambert de Rouvroit, “Aero-thermal performance of a two-dimensional800
highly loaded transonic turbine nozzle guide vane: A test case for inviscid and viscous flow801
computations,” Journal of Turbomachinery 114, 147 (1992).802
47J. C. Tyacke and P. G. Tucker, “Future use of large eddy simulation in aero-engines,”803
Journal of Turbomachinery 137, 081005 (2015).804
48N. Gourdain, F. Sicot, F. Duchaine, and L. Gicquel, “Large eddy simulation of flows805
in industrial compressors: a path from 2015 to 2035,” Philosophical Transactions of the806
Royal Society A 372, 20130323 (2014).807
49X. Yang, J. Sadique, R. Mittal, and C. Meneveau, “Integral wall model for large eddy808
simulations of wall-bounded turbulent flows,” Physics of Fluids 27(2), 025112 (2015).809
50F. Cadieux, J. Sadique, X. I. Yang, C. Meneveau, and R. Mittal, “Wall-modeled large810
eddy simulation of laminar and turbulent separation bubble flows,” in 46th AIAA Fluid811
Dynamics Conference (American Institute of Aeronautics and Astronautics, 2016).812
51L. Cambier, S. Heib, and S. Plot, “The onera elsA cfd software: input from research and813
feedback from industry,” Mechanics & Industry 14, 159–174 (2013).814
52F. Nicoud and F. Ducros, “Subgrid-scale stress modelling based on the square of the815
velocity gradient tensor,” Flow, Turbulence and Combustion 62, 183 (1999).816
53A. Kravchenko and P. Moin, “On the effect of numerical errors in large eddy simulations817
of turbulent flows,” Journal of Computational Physics 131, 310 (1997).818
45
54H. Reichardt, “Complete representation of the turbulent velocity distribution in smooth819
pipes,” Z. Angew. Math. Mech 31, 208 (1951).820
55B. A. Kader, “Temperature and concentration profiles in fully turbulent boundary layers,”821
International Journal of Heat and Mass Transfer 24, 1541 (1981).822
56S. L. Bras, H. Deniau, C. Bogey, and G. Daviller, “Development of compressible large-823
eddy simulations combining high-order schemes and wall modeling,” AIAA Journal 55,824
1152 (2017).825
57T. V. Karman, “Uber laminare und turbulente reibung,” Zeitschrift fur Angewandte Math-826
ematik und Mechanik 1, 233 (1921).827
58K. Pohlhausen, “Zur naherungsweisen integration der differentialgleichung der iaminaren828
grenzschicht,” Zeitschrift fur Angewandte Mathematik und Mechanik 1, 252 (1921).829
59B. Aupoix and S. Viala, “Compressible turbulent boundary layer modelling,” in Transi-830
tional and Turbulent Compressible Flows, Fluids Engineering Division, Vol. 224, edited by831
L. D. Kraland, E. F. Spina, and C. Arakawa (ASME, Hilton Head Island, 1995) p. 139.832
60G. N. Coleman, J. Kim, and R. D. Moser, “A numerical study of turbulent supersonic833
isothermal-wall channel flow,” Journal of Fluid Mechanics 305, 159 (1995).834
61P. G. Huang, G. N. Coleman, and P. Bradshaw, “Compressible turbulent channel flows:835
Dns results and modeling,” Journal of Fluid Mechanics 305, 185 (1995).836
62H. M. Nagib and K. A. Chauhan, “Variations of von karman coefficient in canonical flows,”837
Physics of Fluids 20, 101518 (2008).838
63W. K. George, “Is there a universal log law for turbulent wall-bounded flows?” Philosophi-839
cal Transactions of the Royal Society A: Mathematical, Physical and Engineering Sciences840
365, 789 (2007).841
64T. Houra and Y. Nagano, “Effects of adverse pressure gradient on heat transfer mechanism842
in thermal boundary layer,” International Journal of Heat and Fluid Flow 27, 967 (2006).843
65R. Johnstone, G. N. Coleman, and P. R. Spalart, “The resilience of the logarithmic law844
to pressure gradients: Evidence from dns,” Journal of Fluid Mechanics 643, 163 (2010).845
66J. C. Gillis and J. P. Johnston, “Turbulent boundary-layer flow and structure on a convex846
wall and its redevelopment on a flat wall,” Journal of Fluid Mechanics 135, 123 (1983).847
67N. Afzal, “Analysis of a turbulent boundary layer subjected to a strong adverse pressure848
gradient,” International Journal of Engineering Sciences 21, 563 (1983).849
46
68M. Skote and D. S. Henningson, “Direct numerical simulation of a separated turbulent850
boundary layer,” Journal of Fluid Mechanics 471, 107 (2002).851
69A. Walz, Stromungs-und Temperaturgrenzschichten. Braun (translation in Boundary Lay-852
ers of Flow and Temperature, MIT Press, 1969) (1966).853
70H. Abe, H. Kawamura, and Y. Matsuo, “Surface heat-flux fluctuations in a turbulent854
channel flow up to Reτ = 1020 with Pr = 0.025 and 0.71,” International Journal of Heat855
and Fluid Flow 25, 404 (2004).856
71S. Hoyas and J. Jimenez, “Scaling of the velocity fluctuations in turbulent channels up to857
Reτ = 2003,” Physics of Fluids 18, 011702 (2006).858
72A. Lozano-Duran and J. Jimenez, “Effect of the computational domain on direct simula-859
tions of turbulent channels up to Reτ = 4200,” Physics of Fluids 26, 011702 (2014).860
73M. Lee and R. D. Moser, “Direct numerical simulation of turbulent channel flow up to861
Reτ ' 5200,” Journal of Fluid Mechanics 774, 395 (2015).862
74W. M. Kays, M. E. Crawford, and A. Weigand, Convective Heat and Mass Transfer863
(Mcgraw-Hill (Tx), 2004).864
75D. Modesti and S. Pirozzoli, “Reynolds and mach number effects in compressible turbulent865
channel flow,” International Journal of Heat and Fluid Flow 59, 33 (2016).866
76A. Trettel and J. Larsson, “Mean velocity scaling for compressible wall turbulence with867
heat transfer,” Physics of Fluids 28, 026102 (2016).868
77D. Li, “Revisiting the subgrid-scale prandtl number for large-eddy simulation,” Journal of869
Fluid Mechanics 802, R2 (2016).870
78P. Moin, K. Squires, W. Cabot, and S. Lee, “A dynamic subgrid-scale model for com-871
pressible turbulence and scalar transport,” Physics of Fluids A: Fluid Dynamics 3, 2746872
(1991).873
79C. C. de Wiart, K. Hillewaert, L. Bricteux, and G. Winckelmans, “Implicit les of free874
and wall-bounded turbulent flows based on the discontinuous galerkin/symmetric interior875
penalty method,” International Journal for Numerical Methods in Fluids 78, 335 (2015).876
80O. Cabrit, Modelisation des flux parietaux sur les tuyeres des moteurs a propergol solide,877
Ph.D. thesis (2009).878
81C. Meneveau, T. S. Lund, and W. H. Cabot, “A lagrangian dynamic subgrid-scale model879
of turbulence,” Journal of Fluid Mechanics 319, 353 (1996).880
47
82F. Kremer, Etude numerique d’ecoulements de paroi compressibles: methodes d’integration881
temporelle semi-implicites et application au canal plan turbulent, Ph.D. thesis (2012).882
83C. Cossu and Y. Hwang, “Self-sustaining processes at all scales in wall-bounded turbulent883
shear flows,” Philosophical Transactions of the Royal Society of London A: Mathematical,884
Physical and Engineering Sciences 375 (2017).885
84M. Lagha, J. Kim, J. D. Eldredge, and X. Zhong, “A numerical study of compressible886
turbulent boundary layers,” Physics of Fluids 23, 015106 (2011).887
85P. G. Huang and P. Bradshaw, “Law of the wall for turbulent flows in pressure gradients,”888
AIAA Journal 33, 624 (1995).889
86J. W. Nam and F. S. Lien, “A ghost-cell immersed boundary method for large-eddy sim-890
ulations of compressible turbulent flows,” International Journal of Computational Fluid891
Dynamics 28, 41 (2014).892
87Z. Sun, Y. Zhu, Y. Hu, and S. Zhang, “Direct numerical simulation of a fully developed893
compressible wall turbulence over a wavy wall,” Journal of Turbulence 19, 72 (2018).894
88M. F. Shahab, G. Lehnasch, T. B. Gatski, and P. Comte, “Statistical characteristics of an895
isothermal, supersonic developing boundary layer flow from dns data,” Flow, Turbulence896
and Combustion 86, 369 (2011).897
89Y.-B. Chu, Y.-Q. Zhuang, and X.-Y. Lu, “Effect of wall temperature on hypersonic898
turbulent boundary layer,” Journal of Turbulence 14, 37 (2013).899
90A. Hadjadj, O. Ben-Nasr, M. S. Shadloo, and A. Chaudhuri, “Effect of wall temperature900
in supersonic turbulent boundary layers: A numerical study,” International Journal of901
Heat and Mass Transfer 81, 426 (2015).902
91M. S. Shadloo, A. Hadjadj, and F. Hussain, “Statistical behavior of supersonic turbulent903
boundary layers with heat transfer at M∞ = 2,” International Journal of Heat and Fluid904
Flow 53, 113 (2015).905
92M. S. Shadloo and A. Hadjadj, “Laminar-turbulent transition in supersonic boundary906
layers with surface heat transfer: A numerical study,” Numerical Heat Transfer, Part A:907
Applications 72, 40 (2017).908
48
Newton-Raphson
algorithm
un1 , Tn1
An1, An1T
un , T
n
i An, AnT
Ln , L
nuLn1
, Ln1u
InitialGuess
un1 , Tn1
An1, An1T
i
un1 , Tn1
An1, An1T
i
y1
Mn1x
, Mn1T
x
eun11k
, eun1k
...
Ln1u , Ln1
u , Ln1u2
eun11k
, epn11 ... eun1
1k, epn1
1 ...
Ln1u , Ln1
u , Ln1u2